Alibaba's Qwen 4 reveal was four model names and no spec sheet
Alibaba's Apsara keynote named four Qwen 4 tiers and shipped no spec sheet: no context window, no API id, no price, no release date, no weights, no benchmark.
Vlad Makarovreviewed and published
Alibaba used the opening keynote of its Apsara Conference in Hangzhou on September 22 to put the next generation of Qwen on stage. The names, reported from the room and rounded up in one widely shared write-up: Qwen 4 Max, Qwen 4 Flash, Qwen 4 Plus and Qwen 4 27B. None of the four arrived with a context window, an API identifier, a price per million tokens, a release date, downloadable weights or a benchmark. The official Alibaba Cloud press release says only that Qwen 4 is currently in training. Four names, no spec sheet, plus a firm production date for the chips meant to train what comes next.
The press release is a roadmap, not a product page
Alibaba's announcement is explicit about tense. It says the company "revealed that its next-generation model, Qwen 4, is currently in training," and announced a roadmap for the upcoming Qwen 4.5 and Qwen 5 series, "projected to scale up to 5 to 10 trillion parameters." That is two to four times the current 2.4-trillion-parameter flagship, disclosed as a direction of travel with no architecture, no training-compute figure and no date. The four tier names that traveled out of the room do not appear in the English release at all, which makes them stage reporting rather than published specification.
Eddie Wu, Alibaba Group's chief executive, sold the timeline in units of scale. "Today, the total volume of Machine Thinking is less than 3% of all Human Thinking," he said. "If that volume eventually scales to 1,000x human capacity, the simple math tells us: Machine Thinking still has an enormous growth runway. As machines are becoming the primary force behind Thinking, turning intelligence into a commodity supplied at scale, the truly groundbreaking products of the Machine Intelligence era have not yet arrived." He added that Alibaba Cloud's global data center capacity will surpass 20GW by 2032. Joe Tsai, the group's chairman, called the strategy "guiding AI from technological breakthroughs toward value creation." None of it is a specification.
Four tiers, one blank column
The only usable read on the four names is positional, because no other numbers exist.
| Name shown on stage | Apparent positioning | Published specs |
|---|---|---|
| Qwen 4 Max | Flagship, aimed at the top rival models | None |
| Qwen 4 Flash | Low latency, high volume | None |
| Qwen 4 Plus | Middle tier | None |
| Qwen 4 27B | Smaller, local-class | None |
What does have numbers is what you can buy today
The fair comparison is with the flagship Alibaba actually sells. Qwen3.8-Max, available since August 3, 2026, is a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters, priced at $2 per million input tokens and $6 per million output tokens, with up to a 1M-token context. Qwen 4 inherits none of that in verifiable form. Reuters reported that Alibaba's Hong Kong-listed shares rose about 5% during the session, a re-rating on a roadmap rather than a benchmark.
| Qwen3.8-Max (shipping) | Qwen 4 family | |
|---|---|---|
| Parameter count | 2.4T total, 95B active | not published |
| Input / output price per M | $2 / $6 | not published |
| Context window | up to 1M tokens | not published |
| Availability | since Aug 3, 2026 | in training |
The chips have dates; the models do not
T-Head, Alibaba's chip design unit, announced the Zhenwu V900, an accelerator it says delivers three times the performance of its predecessor, the Zhenwu M890, released in May. These figures are specific:
| Zhenwu V900 spec | Value |
|---|---|
| Performance vs M890 | 3x |
| GPU memory | 216 GB |
| Inter-chip bandwidth | 1,200 GB/s |
| Low-precision support | native FP8 and FP4 |
| Mass production / commercial release | Q1 2027 |
| Zhenwu customer count | over 650 |
A supernode pairing the V900 with the ICN Switch, Panmai SmartNIC and Zhenyue SSD controller can support a cluster of up to 500,000 cards. T-Head also dated next-generation Yitian 720 and Yitian 730 CPUs to 2027, the latter its first on a proprietary microarchitecture and up to 40% better on SPECint2017/GHz than the Yitian 710. So the hardware for the 5-to-10-trillion-parameter models has a quarter attached. The models do not.
The self-improvement figures are the vendor's own
Alibaba's most eye-catching claims concern recursive self-improvement, and all of them are unaudited vendor figures. Over more than a month of fully automated runs spanning pipeline design, data validation, iterative experimentation and error diagnosis, the company says Qwen3.8-Max completed 33 iterative cycles; through autonomous training optimisation and post-training, the updated model raised its Artificial Analysis score from 40 to 45. In one chip-design experiment, Alibaba reports the model self-improved for over 60 hours across the design lifecycle, made more than 10,000 EDA tool calls, produced production-grade chip bus modules and cut chip area by 42% "with zero compromise in performance." A vendor reporting its own scores on its own runs is a claim about the run, not an audit; the 42% figure has no external replication.
The parts that ship, and one clue about what comes next
Not everything announced is a slide. Qwen3.8-LiveTranslate cuts simultaneous-interpretation latency from 2.8 seconds to 2.3 seconds, almost 20%. Qwen-Audio-3.1-TTS-Next builds full cinematic soundscapes from a text script, and the speech suite adds Qwen-Audio-3.1-ASR, TTS and Realtime. Qwen-Image 3.1, promised for later this year, offers native transparent-background generation, a line Alibaba has been building on since Qwen-Image 2.1. The company also introduced a six-tier world-model evaluation framework, a Happyworld Arena benchmarking platform and Qwen Intelligence, a business-facing agentic platform aimed at smartphone makers.
The nearest preview of the next generation is not on the Apsara stage; it shipped in August. Qwen3.8-Flash-Next, open-sourced with roughly 6 billion active parameters, is the closest available signal of where the flagship is heading: about 176 billion total parameters and a 262K native context extensible to 1M. Alibaba has never publicly linked it to Qwen 4. It is a clue, not a confirmation.
Until a number exists, this is a naming exercise
Alibaba has announced a version number before a specification. Qwen 4 Max, Flash, Plus and 27B may turn out to be real products with real capability. What is verifiable today is four strings of text and the word "training." Four things would change that read: a context window, an API identifier, a price and a checkpoint. Until one appears, the defensible conclusion is that this describes a roadmap for a model that does not yet exist, paired with a chip not scheduled for production until 2027.
